reubk/Molmo_7B_Human_Pointing_LoRA

Testing a QLoRA adaptor for allenai/Molmo-7B-D-0924,

Targets attention layer of Transformer backbone and image pooling and projection layers of Vision backbone

Trained on 47 screenshots of a low-poly video game with ragdoll casualties

Evaluated on 44 screenshots of aforementioned video game

Molmo has an edge case where it declares there are no humans in an image:

This custom QLoRA successfully reduces the occurance of these cases

However, pointing to non-human objects is observed to increase.

Comparison of Model performance with and without QLora on Eval dataset

Model	Molmo-7B-D	Molmo-7B-D w/ QLora
Precision	92.1	80.5
Recall	70.4	88.5

Dataset: reubk/RavenfieldDataset